Optimizing Federated Learning based on Weighted Hyperparameter Aggregation in Energy-constrained Building Construction

Federated learning (FL) is compelling for construction safety because sensitive, site-specific data can remain on devices; however, hyperparameter optimization (HPO) in FL is often impractical on resource-constrained edge nodes due to multi-round tuning and communication overhead. Existing FL-HPO approaches (e.g., weight sharing and server-only selection) either incur iterative client-server loops or underutilize client heterogeneity, thereby raising privacy and governance concerns. We propose an asynchronous, one-shot local Bayesian optimization on each client using a small data subset, followed by loss-weighted aggregation of client-proposed hyperparameters on the server to produce a single global set for collaborative training. The pipeline is training-method agnostic and integrated into a Digital Twin (DT) platform for device registration, orchestration, and telemetry. Across public datasets (MNIST, EEGEyeState, Weather) and a private worker-safety dataset, our approach improves convergence and accuracy/MAE relative to simple averaging and reduces response time on Raspberry Pi clients, yielding an average improvement of 6.44% in end-to-end performance while remaining deployable on constrained AIoT hardware. In safety prediction, faster convergence and lower error enable earlier and more reliable hazard indicators, with on-device inference supporting near–real-time alerts to supervisors and wearables. We also discuss scalability, privacy of shared hyperparameters, and task complexity as directions for future work in safety-critical FL.